**8.1 Clinical implications**

This study offered several valuable insights into the prediction challenges in diabetes and similar diseases and explained how they could be tackled. First, throughout this chapter, appropriate machine learning techniques were conducted to model complex interactions among the complications, risk factors and unmeasured factors. For instance, the use of probabilistic graphical models provided a significant improvement in the accuracy of predictive models while reducing uncertainty in disease management. Having adopted DBNs to learn hidden risk factors and effectively understand the AI black box model was the key contribution of this research. The temporal phenotype was identified to represent the overall patterns of disease risk factors for each patient based on the discovered hidden variables over time. The descriptive analytics, in [97], provided valuable insights into the hidden variable effects on stratifying patients into different sub-groups, whether or not they developed the same complications. These findings also explained the influence of the latent variable on the bootstrapped data. Phenotype discovery was utilised to categorise and investigate meaningful subgroups of patients based on how an individual matches historical data. The hybrid type methods in discovering meaningful subgroups and explaining temporal phenotype also led to a better understanding of clinical data as well as aiding to interpret the unmeasured factors while demonstrating their risks.

### **8.2 Future works**

The generalisability of the results presented in this study is subject to certain limitations as follows: This research was conducted to explain and discover the unmeasured factors with a few patients and relatively few features. Thus, this study focused on time-series complex clinical dataset like T2DM, which was a small-sized dataset with an unequal number of patient's follow-up visits (which is common in clinical data). This study was specific to T2DM concept and Bayesian modelling; hence, one fundamental criticism could be the bias towards this dataset and whether the method could be developed in other fields of clinical data in the future. In order to help overcome the limitations discussed in the previous section, the following recommendations are suggested: The originality of the proposed study consisted in its innovative, analytical, and methodological strategies to predict and explain complex clinical data to improve patients' quality of life. A natural progression of this work for a better generalisability should involve extending the latent DBNs model with more hidden variables to capture a greater variety of unmeasured factors to characterise critical changes and produce interesting findings that account more for better explainability and predictability. In addition, to address the limitation related to the small-sized dataset, this work could be extended to further investigation and experimentation into clinical impacts and environmental factors, such as family history, pollution, and glucose. More research also might be conducted to monitor disease progression effectively and detect the underlying patterns of complications, which could provide clinicians with a better understanding of the obtained findings. For example, a greater focus on phenotype discovery could enable assessment of the long-term effects of the temporal phenotype on the patient, which might be done by following qualitative approaches to support the obtained findings from the biomedical literature. The generalisability of the findings obtained in this study might be tested on other data with potentially

*Predicting Type 2 Diabetes Complications and Personalising Patient Using Artificial… DOI: http://dx.doi.org/10.5772/intechopen.94228*

non-stationary, complex, and incomplete data. For instance, the pre-processing approaches, statistical analysis, temporal phenotype, MCI algorithm and the DBNs model could be applied to another complex data (e.g., COVID-19). Therefore, in a new project, a similar patient model to this research was mainly employed, which primarily concentrated on helping healthcare staff in their understanding of how COVID-19 spread and how they could be better prepared.

In the current work as a Post-Doctorate research fellow at Brunel University and University City London (UCL) associated with the BHF Alan Turing Institute jointly funded research project with the collaborators of the project in UCL and GSK. This project aims to develop a computational tool to investigate the action of drug compounds for the treatment of cardiovascular disease and type 2 diabetes which involves: firstly, the construction of a cardiovascular disease (CVD) and Type-II diabetes (T2D) relevant metabolic measures networks, using repeated measures. Secondly, the combination of different causal networks on the same set of metabolic measures. Lastly, the integration to the system of available drug targets and disease information for testing CVD and T2D drugs.

### **Acknowledgements**

test whether the descriptive approaches have contributed to improving the

to model complex interactions among the complications, risk factors and unmeasured factors. For instance, the use of probabilistic graphical models provided a significant improvement in the accuracy of predictive models while reducing uncertainty in disease management. Having adopted DBNs to learn hidden risk factors and effectively understand the AI black box model was the key contribution of this research. The temporal phenotype was identified to represent the overall patterns of disease risk factors for each patient based on the discovered hidden variables over time. The descriptive analytics, in [97], provided valuable insights into the hidden variable effects on stratifying patients into different sub-groups, whether or not they developed the same complications. These findings also explained the influence of the latent variable on the bootstrapped data. Phenotype discovery was utilised to categorise and investigate meaningful subgroups of patients based on how an individual matches historical data. The hybrid type methods in discovering meaningful subgroups and explaining temporal phenotype also led to a better understanding of clinical data as well as aiding to interpret the

unmeasured factors while demonstrating their risks.

This study offered several valuable insights into the prediction challenges in diabetes and similar diseases and explained how they could be tackled. First, throughout this chapter, appropriate machine learning techniques were conducted

The generalisability of the results presented in this study is subject to certain limitations as follows: This research was conducted to explain and discover the unmeasured factors with a few patients and relatively few features. Thus, this study focused on time-series complex clinical dataset like T2DM, which was a small-sized dataset with an unequal number of patient's follow-up visits (which is common in clinical data). This study was specific to T2DM concept and Bayesian modelling; hence, one fundamental criticism could be the bias towards this dataset and

whether the method could be developed in other fields of clinical data in the future. In order to help overcome the limitations discussed in the previous section, the following recommendations are suggested: The originality of the proposed study consisted in its innovative, analytical, and methodological strategies to predict and explain complex clinical data to improve patients' quality of life. A natural progression of this work for a better generalisability should involve extending the latent DBNs model with more hidden variables to capture a greater variety of unmeasured factors to characterise critical changes and produce interesting findings that account more for better explainability and predictability. In addition, to address the limitation related to the small-sized dataset, this work could be extended to further investigation and experimentation into clinical impacts and environmental factors,

such as family history, pollution, and glucose. More research also might be conducted to monitor disease progression effectively and detect the underlying patterns of complications, which could provide clinicians with a better understanding of the obtained findings. For example, a greater focus on phenotype discovery could enable assessment of the long-term effects of the temporal phenotype on the patient, which might be done by following qualitative approaches to support the obtained findings from the biomedical literature. The generalisability of the findings obtained in this study might be tested on other data with potentially

prediction performance of the ordering patterns of complications.

*Type 2 Diabetes - From Pathophysiology to Cyber Systems*

**8.1 Clinical implications**

**8.2 Future works**

**214**

I thank the following individuals for their expertise and assistance throughout all aspects of this study and for their insightful suggestions and careful reading of the manuscript. Dr. Stephen Swift, Dr. Mahir Arzoky and clinicians/scientist at Pavia hospital; especially Lucia Saachi and Luca Chiovato for providing the dataset and their professional advice.
